Drug companies' private protein structures markedly improve AI folding models
A consortium of pharmaceutical companies called the AI Structural Biology (AISB) Network fine-tuned OpenFold3, previously trained only on PDB data, on 20,167 protein–ligand structures from five companies, and on a held-out test set of 1,056 protein–ligand structures the model reached high accuracy on more than half of them, versus about one-third for the public OpenFold3 and around 40% for the open-source model Boltz-2, while also outperforming models trained only on each company's own data, indicating that pooling data is more valuable than keeping it siloed.